Cyber Export Control: A History of Failure from Encryption to Mythos

Last Friday, the White House issued a directive to Anthropic, mandating a restriction on the export of its advanced AI models, Fable and Mythos. Citing unspecified national security concerns, the order prohibited their distribution to individuals outside the United States, as well as to foreign nationals residing within the country. In a swift response, the AI firm immediately discontinued access to both models, rendering them unavailable for the past week.

This incident marks a critical juncture, representing the U.S. government's inaugural significant attempt to employ export controls to manage frontier AI technology. The approach mirrors previous efforts to regulate encryption and spyware, which have yielded inconsistent results. The resolution of this unfolding situation holds substantial implications, not only for Anthropic's access to international markets but also for establishing future regulatory frameworks that other AI laboratories will likely need to navigate.

To provide some context, since its introduction in April, Anthropic has characterized Mythos as a formidable "Doomsday cyber machine" with the potential to severely disrupt the internet if widely disseminated. Consequently, prior to the recent ban, access was meticulously limited to approximately 150 carefully vetted companies and government organizations. The primary objective of this restricted access was to empower digital defenders in securing their software and services, preempting the emergence of similar capabilities among malicious actors.

Reportedly, two distinct events precipitated the ban. The first involved Anthropic granting a South Korean telecommunications company access to Mythos via its limited partner program. U.S. officials expressed alarm upon identifying the company as one they suspected of having ties to China, although the firm, widely reported to be SK Telecom, has since denied any such connection. The second trigger came when Amazon CEO Andy Jassy reportedly informed the administration that Amazon's researchers had discovered a method to circumvent Fable 5's safeguards. Anthropic, however, disputes the characterization of a "jailbreak," describing it instead as a narrow, already-patched vulnerability rather than a fundamental compromise of the model's safety architecture.

Regardless of the differing interpretations, the outcome was immediate: the Commerce Department issued an export control directive. Anthropic, by some accounts, was compelled to swiftly implement access limitations for its products, reportedly within a mere 90 minutes of receiving notification.

Such governmental attempts to regulate potentially dangerous cyber technology are not unprecedented. For decades, governments have utilized export controls to curb proliferation, albeit with a track record that has been, at best, moderate in its success.

A notable historical instance of this approach's failure occurred in the early to mid-1990s, spearheaded by the U.S. government. During this period, computer scientists were innovating encryption technologies to secure data transmission across the burgeoning internet. Among these innovations was Pretty Good Privacy (PGP), a popular software capable of encrypting data, making it virtually impenetrable even if intercepted en route to its intended recipient.

Initially, the U.S. government perceived PGP as a dangerous instrument, fearing it would impede intelligence agencies' ability to monitor email communications. To halt PGP's distribution, the U.S. Customs Service launched a criminal investigation against its creator, Phil Zimmermann, alleging violations of arms export controls. Zimmermann retaliated by publishing PGP's source code as a physical book, an act that ignited what is now famously known as the "Crypto Wars."

Zimmermann ultimately prevailed in a pivotal victory when the investigation was closed. This outcome paved the way for the development and widespread adoption of crucial end-to-end encryption algorithms, now utilized by billions of users on platforms like Signal and WhatsApp.

Moving into the early 2010s, researchers uncovered instances of Western-made spyware being deployed against dissidents in the Middle East. This discovery prompted several governments to expand the Wassenaar Arrangement, an international treaty designed to restrict the export of "dual-use" software and technologies—those with both civilian and military applications.

The core objective of this expansion was to categorize surveillance and hacking software as dual-use, thereby mandating that spyware manufacturers obtain export licenses for selling their products internationally.

However, the Wassenaar Arrangement has always been hampered by two inherent weaknesses. Firstly, several nations do not adhere to the agreement, including Israel, which is home to some of the world's most active spyware developers.

Secondly, the agreement relies on member states to apply its provisions to companies within their borders at their own discretion. For instance, the Italian government, for a period, granted Hacking Team, then a prominent spyware maker, a license to export its tools globally, despite the company's documented history of selling spyware to oppressive regimes that subsequently used it to target journalists and human rights activists.

Since then, other European countries have demonstrated similar leniency towards spyware manufacturers. Despite numerous scandals, Europe, a hub for many spyware and hacking tool developers, has consistently struggled to curb the export of spyware to authoritarian governments. Critics contend that a recent renewed initiative among the 27-member bloc to address its escalating problem of spyware exports to authoritarian states "does not go far enough."

Furthermore, several spyware companies, such as Intellexa, a sanctioned consortium of spyware firms, have simply relocated their operations to countries with more relaxed export controls. Other developers have similarly sought to establish operations in places like Saudi Arabia for comparable reasons.

Despite these challenges, there have been some notable successes. The Germany-based spyware maker FinFisher, for example, ceased operations in 2022 following a multi-year investigation by German prosecutors. The probe focused on allegations of the company selling spyware to Turkey without the requisite export license, with investigators previously confirming FinFisher spyware had been deployed on the phones of critics of the Turkish government.

As of this writing, the standoff between Anthropic and the administration persists. There remains a plausible scenario where the administration may relent and lift the restriction to maintain the global competitiveness of American AI companies. Such a move would tacitly acknowledge that AI labs worldwide, including those in China, are likely to achieve comparable capabilities irrespective of U.S. export limitations. Alternatively, American AI companies could face the prospect of requiring government approval for all foreign customer engagements, a compliance burden that would undoubtedly impact their profitability.

Considering the historical experiences of governments attempting to control software proliferation, it appears that government-mandated export controls are unlikely to be the definitive solution for preventing malicious actors from misusing powerful, dual-use cyber technologies.